This paper introduces a new computer-vision method (COCOS) to automatically recognize contrails in satellite images captured by the MSG SEVIRI instrument. In contrast to many recently introduced methods for the detection of contrails in images, it does not explicitly consist of neural networks. This approach is motivated by the idea that it leads to an interpretable, adaptable detection algorithm that can be adjusted to new satellite instruments in the near-future. The improvement in performance, as evaluated using a dataset consisting of manually annotated satellite images, is considerable when compared to previous methods for contrail detection in MSG SEVIRI images

The paper is certainly within the scope of AMT, it is very well-written and it was an enjoyable read: I have only few line-by-line comments, which are given in a separate section below. I do have some more general comments. Once these are considered by the authors, I think the paper is ready to be published.

My first comment concerns section 2.4 “Contrail Ground Truth” which describes the dataset of annotated MSG images used for developing and evaluating the detection algorithm. After reading this section, I still have many remaining questions:

• Broadly speaking, how were labelers instructed to identify contrails? Was this in any way similar to the instructions used by (Meijer et al., 2022) and (Ng et al., 2024)?
• How were scenes selected for labeling, and what is the resulting spatio-temporal distribution of annotated scenes?
• What information (other than the image itself) was presented to the labelers? For example, both previously mentioned annotated datasets provided labelers with a sequence of images leading up to the relevant image.

For my second comment, I refer to a passage from the introduction:

By avoiding machine learning methods, the authors aim to facilitate the future adaptation of COCOS to the Meteosat Third Generation satellite. Furthermore, we gain a deeper understanding of the factors affecting contrail detection—such as why a contrail is detected or missed, which spectral channels contribute most, and which contrail characteristics are most informative, thereby strengthening the use of COCOS for assessing temporal and spatial contrail distributions, evaluating the effectiveness of mitigation strategies (such as flight rerouting), and improving the representation of aviation-induced clouds in climate models and contrail models.

I find the above motivation to avoid a machine-learning based approach sensible. However, I think that the terminology “avoiding machine learning methods” should be modified slightly, as the COCOS algorithm introduced in the paper still relies on the output of another machine learning approach: CiPS. Secondly, I think that the paper’s contribution to a “Deeper understanding of the factors affecting contrail detection” could be expanded. Although the analysis of the algorithm’s performance (and its dependency on background/contrail properties) is clear, I would be very interested in learning what parts of /tests within the algorithm contribute mostly to the improvement in performance compared to previous MSG SEVIRI algorithms. For example, suppose the CiPS output is a very important input to COCOS: this knowledge could be used in the design of new contrail detection approaches.

Line-by-line comments

• Line 26, it may of interest to the authors to also cite (Sonabend-W et al., 2025)
• Line 87, “This also ensures that contrail detection remains independent of the time of the day” Would it perhaps be better to state this as “remains possible throughout the day”, since variations in scene properties may still occur throughout the day which could possibly affect contrail detection performance?
• Line 89: there seems to be a “8” missing here
• Line 116: Missing a space between “contrails” and “(“
• Line 222: Note that “gaussian weighted” is spelled differently here than for example line 220
• Line 415: Seems like there is an additional space at the start of this sentence.
• Figure 8: I would strongly consider replacing the pie charts with other means of visualization. For example, the distribution of “linearity” and “ice optical thickness” could be visualized with CDFs/PDFs.

References

Meijer, V.R., Kulik, L., Eastham, S.D., Allroggen, F., Speth, R.L., Karaman, S., Barrett, S.R.H., 2022. Contrail coverage over the United States before and during the COVID-19 pandemic. Environ. Res. Lett. 17, 034039. https://doi.org/10.1088/1748-9326/ac26f0

Ng, J.Y.-H., McCloskey, K., Cui, J., Meijer, V.R., Brand, E., Sarna, A., Goyal, N., Van Arsdale, C., Geraedts, S., 2024. Contrail Detection on GOES-16 ABI With the OpenContrails Dataset. IEEE Trans. Geosci. Remote Sens. 62, 1–14. https://doi.org/10.1109/TGRS.2023.3345226

Sonabend-W, A., Geraedts, S., Goyal, N., Ng, J.Y.-H., Van Arsdale, C., McCloskey, K., 2025. Observing long-lived longwave contrail forcing. EGUsphere 1–25. https://doi.org/10.5194/egusphere-2025-3739

I read carefully through this interesting article. I have some suggestions. 

1. Contrails form at temperatures below about -40C, unless there are some extenuating circumstances. "Hole Punch Clouds". I suggest using temperature as another way of checking your results.

2. It would be interesting to use ERA5 relative humidity/temperature data to see how well those fit with your analysis

3. EarthCare lidar data will be helpful in the future to further quantify how well your method fits the observations. You did use the CALIOP data in quantifying your results and the EarthCare lidar data can be used in the future.